以多核心圖形處理器加速影像處理之研究

Abstract

本論文研究以多核心圖形處理器(Multicore Graphic Processing Units)加速影像處理演算法，我們以全向圖(omnidirectional pictures)轉換成全景圖(panoramic pictures)及車牌辨識(vehicle license plate recognition)系統為例，提出平行演算法並以多核心圖形處理器進行相關演算法加速。 論文首先針對橢圓拋物曲面全向圖轉換成全景圖的演算法進行平行化研究，本論文提出了一個階層式的平行架構包含資料平行(data parallelism)與任務平行(task parallelism)兩個階層，其中資料平行階層是透過執行圖形處理器的大量執行緒平行轉換每個像素從全向圖移轉至全景圖，而任務平行階層是透過圖形處理器多串流技術(multiple stream)，以管線化(pipelining)的方式平行執行多個影像的轉換。任務平行可以藉由重疊影像處理器的核心運算與資料傳輸的執行時間來改善整體的效能。實驗結果顯示相較於CPU，透過圖形處理器，我們可以得到6.33倍的改善。 論文第二部分，我們針對車牌辨識系統進行平行化研究，一個車牌辨識系統主要包含車牌定位、車牌校正、文字切割與文字辨識等四大步驟。首先在車牌定位部分，我們透過灰階轉換、直方圖等化、二值化、輪廓萃取與剛性物體偵測之核心演算法取得車牌的位置，然後在車牌校正方面，我們使用仿射轉換中的單映性以校正歪斜的車牌。在文字分割方面，我們利用輪廓萃取及邊緣偵測將文字與車牌面積進行計算，並將車牌中的文字分割取出。最後在文字辨識部份，我們利用樣板比對法(template matching)作為文字辨識的方法，為了縮短辨識系統計算的時間，我們透過圖形處理器加速車牌文字辨識的計算速度相較於CPU，我們可以得到100倍的改善。

關鍵字：多核心圖形處理器、影像處理、全景圖轉換、車牌定位、車牌辨識This thesis proposes to accelerate image processing algorithms using multicore Graphic Processing Units (GPUs). Taking the transformation of omnidirectional pictures to panoramic pictures and vehicle license plate recognition system as cases, we propose parallel approaches to accelerate relative image processing algorithms using GPUs. First, we study to parallelize the transformation of elliptical omnidirectional pictures to panoramic pictures. We propose a hierarchical parallelism architecture which includes data parallelism and task parallelism. The data parallelism issues large amount of threads to simultaneously map each pixel of an elliptical omnidirectional pictures to the corresponding position in a panoramic pictures. On the other hand, the task parallelism adopts multiple stream technique to pipeline the transformation of multiple images. The task parallelism improves the overall throughput by overlapping the latency of kernel computation and data transmission time. Experimental results demonstrate that the proposed algorithm achieves 6.33 times of performance improvement as compared to CPU counterpart. Furthermore, we study on the parallelization of vehicle license plate recognition system. A vehicle license plate recognition system composes of four stages including plate localization, plate calibration, text segmentation, and text recognition. First, in the step of plate localization, we obtain the position of a plate via the steps of gray transformation, histogram equalization, image binarization, contour extraction, and rigid object detection. Then, in the step of plate calibration, we adopt single affine transformation to calibrate skew license plates. Furthermore, in the step of text segmentation, we segment texts by extracting the edges and contours of texts and compare their area with that of a license plate. Finally, we perform text recognition using template matching algorithm. In order to reduce the elapsed time of text recognition, we propose to accelerate template matching algorithm using GPUs, compared to the CPU, we can get 100 times improvement.

Keywords: multicore graphic processing units, image processing, panoramic pictures transformation, vehicle license plate localization, vehicle license plate recognition